perf(ui): precalculate document embeddings for graph similarity

[ishaanxgupta]

Summary:

• Pre-calculates Array.from(doc.summaryEmbedding) out of the nested for loop in useGraphData.
• This converts $O(n^2)$ allocations to $O(n)$, significantly reducing memory GC cycles and improving rendering frame times on large graph datasets.

In packages/ui/memory-graph/hooks/use-graph-data.ts, we were calling Array.from(doc.summaryEmbedding) repeatedly within an O(n^2) inner loop. This triggered heavy garbage collection and continuous array allocation when iterating over document pairs for semantic similarity.
I pulled the embedding extraction into an O(n) mapping pass before the nested loops:

// Precompute embeddings to avoid Array.from in the O(n^2) inner loop
const docEmbeddings = filteredDocuments.map((doc) =>
    doc?.summaryEmbedding ? Array.from(doc.summaryEmbedding) : null,
);

Then we simply reference docEmbeddings[i] and docEmbeddings[j] within the $O(n^2)$ loop. This eliminates the heavy GC pressure and directly improves framing times during document selection and graph panning operations.